This repository implements IRAT which is the codes of the paper "Individual Reward Assisted Multi-Agent Reinforcement Learning" (https://proceedings.mlr.press/v162/wang22ao/wang22ao.pdf).
This codes are implemented based on the repository https://github.com/marlbenchmark/on-policy.
This repository contains the experiments in MPE (continue and discrete), Multiwalker (SISL) and StarCraftII (SC2.4.10).
There are two ways to conduct an experiment and train.
-
The bash scripts of these experiments are presented in directory "irat_code/scripts". We can run these bash scripts directly to start an experiments.
-
We also can run the files in directory"irat_code/scripts/train" using python3.
Adding or changing the parameters in the .sh files or in command line will conduct different experiments and get different results. The parameters in .sh files are default parameters used in experiments of paper "Individual Reward Assisted Multi-Agent Reinforcement Learning".
If you find this repository useful, please cite our paper:
@inproceedings{DBLP:conf/icml/WangZHWZGHLF22,
author = {Li Wang and
Yupeng Zhang and
Yujing Hu and
Weixun Wang and
Chongjie Zhang and
Yang Gao and
Jianye Hao and
Tangjie Lv and
Changjie Fan},
editor = {Kamalika Chaudhuri and
Stefanie Jegelka and
Le Song and
Csaba Szepesv{\'{a}}ri and
Gang Niu and
Sivan Sabato},
title = {Individual Reward Assisted Multi-Agent Reinforcement Learning},
booktitle = {International Conference on Machine Learning, {ICML} 2022, 17-23 July
2022, Baltimore, Maryland, {USA}},
series = {Proceedings of Machine Learning Research},
volume = {162},
pages = {23417--23432},
publisher = {{PMLR}},
year = {2022},
url = {https://proceedings.mlr.press/v162/wang22ao.html}
}